Japanese TTS reading kanji wrong: names, numbers and how to force a reading

Why Japanese text to speech misreads kanji

Most kanji have at least two kinds of reading. On'yomi descend from the Chinese pronunciation of the character, and kun'yomi are native Japanese words attached to it. Wikipedia's article on kanji notes that "some common kanji have ten or more possible readings", and gives 生 as an example with eight basic ones (Wikipedia: Kanji). Some words go further: 今日 is read きょう as a whole word, not from its characters, a pattern called jukujikun.

A reader picks the right reading from context without noticing. A text-to-speech engine has to commit to one reading before it produces any sound. When the sentence is short, as on a phone call or in a subtitle line, there is less context to go on, and the engine is more likely to guess wrong.

Japanese writing has a tool for this, which shows how real the problem is even for people. Furigana, the small kana printed beside kanji, exist to show readings a reader could not otherwise know. Forms carry a separate field for the reading of a name, labelled フリガナ, because the kanji alone does not settle it. Hebrew and Arabic text to speech face a similar guessing problem with unwritten vowels, as our guide to Hebrew TTS without niqqud describes.

Where misreadings cluster: names, dates and everyday words

Errors are not spread evenly through a script. They cluster in a few kinds of word, and knowing which ones lets you check a long script quickly instead of listening to every line with equal suspicion.

How to force a reading

There are two ways to tell an engine which reading you mean: markup, where the engine supports it, or the text itself, which works everywhere. Which you use depends on the engine, not on preference.

Why converting whole sentences to kana backfires

If one kana word fixes a reading, it is tempting to remove kanji altogether. Bunpro, a grammar-study site, tried exactly that in May 2024 (Bunpro community). Staff said they had programmatically removed the kanji from their example sentences so the correct kana would be read every time. The bulk run went wrong in a way their small test had not shown: in their words, "the audio can't seem to tell where one word ends and another begins". Users reported misread particles and unnatural phrasing, and staff said they were looking at rolling back to the old audio.

The reason is that kanji carry information beyond the reading. Japanese has no spaces, so the switch between kanji and kana is one of the main signals of where a word starts. Kanji also separate words that sound alike: はし is 箸 (chopsticks), 橋 (bridge) or 端 (edge), and in Tokyo Japanese each has a different pitch pattern. Write はし in kana and you have removed the clue a voice needs to choose the pitch.

So keep kana replacements rare and local. Replace the word, not the phrase, and prefer the form a Japanese writer would accept, such as katakana for a name read back to its owner. If many words in a script need kana, rewrite the sentence with a less ambiguous word instead.

Pitch accent, intonation and subtitles

Correct readings are not the whole job. Tokyo Japanese distinguishes some words only by pitch accent, the pattern of high and low pitch across a word: 端 is flat, 箸 is high on the first syllable and 橋 on the last (Wikipedia: Japanese pitch accent). Ordinary writing never marks it. Commercial narration producers treat it as a separate problem from readings. When Dai Nippon Printing announced an AI narration system in June 2021, it named both misread kanji and the different intonation of 橋, 箸 and 端 as targets. It claimed about 50 to 70 percent fewer misreadings than its earlier synthesis work (Impress Watch).

Markup helps on some engines. Polly's pronunciation-kana alphabet can encode the pitch drop in a word with an apostrophe. Pāṇini has no pitch or intonation control; the settings it exposes are the voice, the speed (0.5 to 2.0) and the output format. What you can control is the text: shorter sentences, a 、 where a native speaker would breathe, and one subtitle cue per sentence so each clip starts and ends cleanly.

For subtitles, intonation across the whole phrase matters as much as any single reading, because a viewer hears each line on its own, timed against the picture. Render each line separately, play it against the cue, and have a native speaker mark any line whose stress sounds wrong. Rewording a line almost always fixes it faster than fighting the engine.

Checking readings before you publish or go live

Most reading errors can be caught before a listener hears them, because they come from a predictable set of words. Start by exporting every name, date, time, amount and product term your script or agent will say. Then have a native speaker flag each one that could be read two ways in context, with the intended reading beside it.

Replace only the flagged words with kana, in your own code, before the text reaches the voice. For AI voice agents that read back customer names, store each name with its reading, the same フリガナ field a Japanese form collects, and send the kana to the voice. Never let the agent read a name straight from a kanji transcript; ask the caller to confirm it, or match it to a record that holds the reading.

Finally, render every flagged sentence and have the native speaker listen, on a phone line for calls or against the cue for subtitles. Repeat the check after any change to the script or the voice, since a new model version can resolve the same word differently. Our guide to testing a multilingual voice agent covers the wider process.